Off-line and on-line combined multi-period traffic estimated value determination method and device

Abstract

The invention relates to an offline and online combined multi-period traffic estimated value determination method and device. The method comprises the following steps: acquiring a main output header result of target traffic in a traffic set; selecting a sub-output head result matched with the main output head result in a corresponding time period from a data pool according to the flow generation attribute and the exposure duration of the target flow; determining an addition coefficient according to the main output header result and the sub output header result, wherein the addition coefficient is used for indicating the correlation degree of the full-cycle conversion rate of the target flow and the conversion rate in the corresponding time period; calibrating the current conversion number of the target traffic based on the addition coefficient to obtain an addition conversion number of the target traffic; and determining a flow decision estimated value at the current moment according to the sum of the addition conversion numbers of all the flows at the current moment and the main output head result of each flow. According to the invention, the accuracy of the traffic conversion estimated value can be improved.

Description

Technical Field

[0001] The present application relates to the field of advertising delivery technology, and in particular to a method and device for determining multi-period traffic estimation values ​​by combining offline and online methods. Background Art

[0002] In areas such as advertising and traffic conversion prediction, the accuracy of model predictions directly impacts business decisions. In practice, to balance inference efficiency and prediction comprehensiveness, models with multiple prediction results are often used. These models generate multiple predictions for the same traffic. The primary output (such as the full-cycle conversion rate) must meet online real-time inference requirements and be used directly for ad ranking. Sub-outputs (such as time-based conversion rates) are primarily used for offline calibration tasks due to their longer inference times and are not used in real-time ranking.

[0003] However, when calibrating the estimated results of these models, there is a significant gap between actual conversion behavior and the estimated results after traffic exposure. This high-latency conversion leads to calibration errors, which is a core challenge. Existing calibration methods generally use the following two approaches.

[0004] Forcibly lowering estimates to match current conversions: To ensure that estimates match current conversions, they are forcibly lowered. (For example, six hours after an ad is exposed, the model estimates a 24-hour conversion rate of 5%, but only 2% of conversions are actually achieved. Traditional methods calibrate the estimate to 2%). This practice ignores the fact that delayed conversions haven't yet occurred, resulting in the calibrated estimate being lower than the actual full-cycle conversion rate. This, in turn, causes the ad to be undervalued in the auction due to the low estimated value.

[0005] Leveraging historical delayed conversion patterns to compensate for future conversions: This approach uses historical data to identify delayed conversion patterns (e.g., a 30% delay in conversions after similar ad exposure in the past) and simply adds the current conversion number to the historical delay percentage to compensate for possible future conversions. This approach overly relies on historical patterns, is insensitive to real-time traffic fluctuations, and impacts decision-making accuracy.

[0006] Therefore, the existing technology may have inaccurate conversion estimates when estimating traffic. Summary of the Invention

[0007] The present application provides a method and device for determining multi-period traffic estimate values ​​by combining offline and online methods to solve the problem of inaccurate traffic conversion estimate values.

[0008] In a first aspect, the present application provides a method for determining a multi-period traffic estimate value by combining offline and online methods, the method comprising:

[0009] Obtaining a main output header result of a target flow in the flow set, wherein the main output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time reasoning;

[0010] Selecting a sub-output header result that matches the main output header result within a corresponding time period from the data pool based on the traffic generation attributes and exposure duration of the target traffic, wherein the sub-output header result is the estimated conversion rate of the target traffic obtained through offline inference. The data pool stores the offline inference results of each traffic flow in different time periods;

[0011] Determining an addition coefficient based on the main output head result and the sub-output head result, wherein the addition coefficient is used to indicate the degree of correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period;

[0012] Calibrate the current conversion number of the target flow based on the addition coefficient to obtain the addition conversion number of the target flow;

[0013] The estimated value of the traffic decision at the current moment is determined based on the sum of the added conversion numbers of all traffic at the current moment and the main output head results of each traffic.

[0014] Optionally, selecting a sub-output header result that matches the main output header result within a corresponding time period from a data pool according to the traffic generation attribute and exposure duration of the target traffic includes:

[0015] According to the traffic generation attribute of the target traffic, selecting sub-output header results of multiple time periods that match the main output header result from the data pool;

[0016] Determine a corresponding time period of the exposure duration of the target traffic, wherein the corresponding time period is used to indicate that the traffic exposure duration is greater than m and less than n, where m and n are positive integers;

[0017] Determine a sub-output header result of the target flow in the corresponding time period, wherein the sub-output header result is a cumulative estimated conversion rate within a time period from 0 to n.

[0018] Optionally, selecting, from a data pool, sub-output header results of multiple time periods that match the main output header result according to the traffic generation attribute of the target traffic includes:

[0019] According to the divided time slices of the target traffic and the model name and model version of the adopted inference model, sub-output header results of multiple time periods that match the main output header result are selected from the data pool.

[0020] Optionally, before selecting, from the data pool according to the traffic generation attribute and exposure duration of the target traffic, a sub-output header result that matches the main output header result within a corresponding time period, the method further includes:

[0021] Export the traffic after online inference is completed by time slices, and divide the traffic by the model name and model version of the inference model used by the traffic;

[0022] Perform offline inference on the divided traffic according to the traffic usage inference model, and output sub-output header results for multiple time periods;

[0023] The sub-output header results of the multiple time periods are stored in a data pool.

[0024] Optionally, the traffic exported by the shards is different batches of traffic, and each batch of traffic sequentially executes the processes of data partitioning, data inference, and data storage. The process steps of different batches of traffic are cross-parallel at the same time point.

[0025] Optionally, the method further includes:

[0026] If the main output header result cannot be matched to the sub-output header result, the addition coefficient is set to 1.

[0027] Optionally, determining the traffic decision estimate at the current moment based on the sum of the added conversion numbers of all traffic at the current moment and the main output header result of each traffic includes:

[0028] Calculate the sum of the additive conversion numbers of all flows at the current moment and the difference between the sum of the main output head results of each flow;

[0029] Based on the difference, adjusting the sum of the main output head results of each flow rate by a preset calibration method so that the sum of the adjusted main output head results matches the sum of the additive conversion numbers of all flow rates at the current moment;

[0030] The sum of the main output header results after integration is used as the estimated flow decision value at the current moment.

[0031] In a second aspect, the present application provides an offline and online combined multi-period flow rate estimation determination device, the device comprising:

[0032] An acquisition module is used to obtain a main output header result of a target flow in a flow set, wherein the main output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time reasoning;

[0033] a selection module, configured to select, from a data pool, a sub-output header result that matches the main output header result within a corresponding time period based on the traffic generation attributes and exposure duration of the target traffic, wherein the sub-output header result is an estimated conversion rate of the target traffic obtained through offline inference, and the data pool stores offline inference results for each traffic flow in different time periods;

[0034] A first determining module is configured to determine a bonus coefficient based on the main output header result and the sub-output header result, wherein the bonus coefficient is used to indicate a correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period;

[0035] a calibration module, configured to calibrate the current conversion number of the target flow based on the addition coefficient to obtain the addition conversion number of the target flow;

[0036] The second determination module is used to determine the traffic decision estimated value at the current moment according to the sum of the added conversion numbers of all flows at the current moment and the main output header result of each flow.

[0037] In a third aspect, the present application provides an electronic device comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus.

[0038] In a fourth aspect, the present application further provides a computer storage medium storing computer executable instructions, wherein the computer executable instructions are used to execute the offline and online combined multi-period flow estimate determination method described in any one of the above items of the present application.

[0039] The above technical solution provided by the embodiment of the present application has the following advantages compared with the prior art: first, the full-cycle estimated conversion rate (main output header result) of the target traffic through online real-time reasoning is obtained to provide a long-term benchmark for calibration; at the same time, offline reasoning is performed on the traffic to generate a time-segment cumulative estimated conversion rate (sub-output header result) and store it in the data pool. By matching the traffic generation attributes and exposure duration, the sub-output header results of the corresponding time period are matched from the data pool, and the additive coefficient is obtained by the ratio of the main output header result and the sub-output header result to quantify the degree of correlation between the full-cycle conversion and the current time period conversion. The current conversion number is then calibrated using the coefficient to obtain an additive conversion number that integrates short-term actual conversion and long-term expectation. Finally, the sum of the additive conversion numbers of all traffic is summarized, and the traffic decision estimate is obtained by recalibrating it in combination with the sum of the main output header results. In this process, the offline reasoned time-segment data and the online real-time full-cycle data complement each other, which not only ensures the timeliness of calibration, but also improves the accuracy of the conversion estimate through multi-dimensional matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0042] One or more embodiments are exemplarily illustrated by pictures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements. Unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0043] Figure 1 A flowchart of a method for determining multi-period flow rate estimates by combining offline and online methods provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of the complete process of a method for determining multi-period traffic estimate values ​​by combining offline and online methods provided in an embodiment of the present application;

[0045] Figure 3 A schematic diagram of the structure of an offline and online combined multi-period flow rate estimation determination device provided in an embodiment of the present application;

[0046] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0048] The disclosure below provides many different embodiments or examples for implementing different structures of the present application. In order to simplify the disclosure of the present application, the components and settings of specific examples are described below. Of course, these are merely examples and are not intended to limit the present application. In addition, the present application may repeat reference numbers and / or letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0049] In order to solve the problem of inaccurate traffic conversion estimates mentioned in the background technology, the embodiment of the present application generates main output head results and sub-output head results through offline and online combination, calculates the addition coefficient by matching the sub-output head results according to the exposure duration, thereby calibrating the current conversion number and solving the problem of delayed conversion and underestimation of the value of short-term unconverted traffic.

[0050] The application scenarios of the embodiments of the present application include but are not limited to: advertising on e-commerce platforms, traffic distribution on content platforms, etc.

[0051] The following will be combined with specific implementation methods to provide a detailed description of a method for determining a multi-period flow estimate value by combining offline and online methods provided by the embodiment of the present application, taking application to a server as an example. Figure 1 The specific steps are as follows:

[0052] Step 101: Obtain a main output header result of a target flow in a flow set, wherein the main output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time inference;

[0053] Step 102: Based on the traffic generation attributes and exposure duration of the target traffic, a sub-output header result that matches the main output header result within the corresponding time period is selected from the data pool. The sub-output header result is the estimated conversion rate of the target traffic obtained through offline inference. The data pool stores the offline inference results of each traffic flow in different time periods.

[0054] Step 103: Determine a bonus coefficient based on the main output header result and the sub-output header result, wherein the bonus coefficient is used to indicate the degree of correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period;

[0055] Step 104: calibrating the current conversion number of the target flow rate based on the addition coefficient to obtain the added conversion number of the target flow rate;

[0056] Step 105: Determine the traffic decision estimate value at the current moment based on the sum of the added conversion numbers of all traffic at the current moment and the main output header result of each traffic.

[0057] First, the terms involved in this application are explained, including the following.

[0058] Traffic collection: refers to all the traffic to be evaluated generated by the platform within a certain period of time (such as product click traffic on e-commerce platforms and video viewing traffic on content platforms), which is the object pool for overall evaluation.

[0059] Target traffic: A single specific traffic selected from the traffic set (such as the traffic generated by user A clicking on product B) is the basic unit of individual evaluation.

[0060] Main output header result: refers to the full-cycle estimated conversion rate (such as the total conversion probability within 7 days after exposure) obtained through online real-time inference of the target traffic, which is used to quickly reflect the long-term conversion potential of the traffic.

[0061] Sub-output header results: Estimated conversion rate by time period (such as 0-6h, 0-12h, 0-24h, etc.), used to reflect the conversion expectations of target traffic in different time periods and assist in calibrating the main output header results.

[0062] Traffic generation attributes: Attributes used to associate the main output header data with the sub-output header data. These attributes typically include the same time slice (e.g., the same 5-minute period) and the same model version (e.g., an algorithm version with consistent inference logic), ensuring that both belong to the same batch of traffic and use the same inference logic.

[0063] Add-on coefficient: The ratio of the main output header result (full-cycle conversion rate) to the sub-output header result (time-divided conversion rate), which is used to quantify the proportion of long-term delayed conversions to short-term conversions in the full-cycle conversion, reflecting the degree of correlation between long-term and short-term conversions.

[0064] Current conversions: The actual number of conversions that have occurred from the time the target traffic was exposed to the current moment (such as the number of purchases after clicks, the number of likes after video viewing), which is a short-term result that can be observed in real time.

[0065] Added conversion number: The value obtained by amplifying the current conversion number by the addition coefficient is used to estimate the potential conversion scale of the traffic over the entire cycle, integrating short-term actual conversions with long-term expectations.

[0066] Traffic decision estimate: The final estimate obtained by integrating the calibration results of all traffic at the current moment is used to guide business decisions such as traffic allocation and prioritization (such as allocating resources to high-potential traffic).

[0067] In step 101, the server selects any specific flow from the traffic set consisting of all active traffic on the platform as the target flow. The target flow can be traffic generated by a user clicking on an ad, browsing a product, or other actions. The server extracts the main output header data corresponding to the target flow. The main output header data contains basic information about the target flow, such as user characteristics and exposure time, as well as key results obtained through online real-time inference. The server obtains the main output header result from the main output header data. The main output header result is the estimated conversion rate of the target flow over the entire cycle. The full-cycle estimated conversion rate is the probability of the flow converting (e.g., purchasing a product, clicking a link, etc.) within a preset complete cycle (e.g., 24 hours, 7 days, etc.), starting from the moment the flow is exposed. This step leverages the online low-latency real-time inference mechanism to generate the full-cycle conversion expectation of the target flow at the moment it is generated, providing a basic full-cycle benchmark value for the subsequent calibration process. In this way, the long-term conversion potential of traffic can be quickly captured during its real-time flow, meeting the real-time requirements of dynamic advertising decision-making and avoiding missing the optimal decision opportunity due to waiting for data to be obtained after the full cycle has ended.

[0068] For example, in the traffic set of the e-commerce platform, the traffic generated by the user clicking on a mobile phone advertisement is selected as the target traffic. The main output header result obtained from the main output header data of the traffic is an estimated conversion rate of 8% for the entire 24-hour cycle, that is, the estimated probability that the user will purchase this mobile phone within 24 hours after clicking the advertisement is 8%.

[0069] In step 102, the server first determines the traffic generation attributes of the target traffic, which include the time slice, model name, and model version. It then determines the exposure duration of the target traffic, that is, the time elapsed from the time the traffic was seen (exposed) by the user to the current moment. Based on this information, the server retrieves matching data from the data pool: the data pool pre-stores the time-segment estimation results (such as 0-6 hours, 6-12 hours, and 12-24 hours) generated by offline inference for each traffic flow. The server selects sub-output header results that are consistent with the target traffic generation attributes and whose time period covers the exposure duration (such as an exposure duration of 8 hours corresponding to a 6-12 hour period result). This step ensures a strong correlation between the sub-output header results and the target traffic through "attribute matching + duration adaptation", providing a time-segment benchmark that fits the current conversion stage for the addition coefficient calculation, avoiding calibration bias caused by cross-attribute data.

[0070] For example, the generation attribute of the target traffic is "14:30-14:35 time slice + model v3.0", the exposure time is 8 hours, and the sub-output header result matched to the 6-12h period from the data pool is 7%, that is, the estimated conversion rate of the traffic within 6-12 hours after exposure is 7%.

[0071] In step 103, the server uses the main output header result obtained in step 101, that is, the full-cycle estimated conversion rate of the target traffic, and the sub-output header result of the corresponding time period selected in step 102, that is, the estimated conversion rate of the target traffic in the corresponding time period, to calculate the ratio of the two to obtain the addition coefficient. The addition coefficient can quantify the degree of correlation between the full-cycle conversion potential of the target traffic and the conversion progress of the current time period, thereby establishing a quantitative correlation between short-term conversion and long-term conversion. When the addition coefficient is greater than 1, it means that the full-cycle conversion expectation is higher than the current time period performance (there is room for delayed conversion); when the addition coefficient = 1, it indicates that the current time period performance is consistent with the full-cycle expectation. This correlation provides a basis for subsequent calibration work, so that short-term conversion data can be reasonably mapped to the full-cycle dimension, effectively solving the problem of one-sided estimation caused by relying solely on current data in high-delay conversion scenarios.

[0072] For example, the result of the main output head obtained in step 101 is an estimated conversion rate of 8% for the entire 24-hour cycle, and the result of the 0-6 hour sub-output head corresponding to a 5-hour exposure time obtained in step 102 is 5%, then the addition coefficient is 8% ÷ 5% = 1.6.

[0073] In step 104, the server counts the current conversion number of the target traffic from the exposure moment to the current moment. The current conversion number refers to the number of times the target traffic is actually converted during this period. The current conversion number is multiplied by the additive coefficient to calibrate the current conversion number, and the result is the additive conversion number of the target traffic. This process amplifies the current conversion number to the full-cycle dimension through the additive coefficient, so that the amplified additive conversion number includes both the actual conversion situation that has occurred and the conversion expectations for the time period that has not yet occurred, which can objectively reflect the real conversion potential of the target traffic in the full cycle. This step effectively avoids the problem of underestimating the value of traffic due to delays in conversion, resulting in a low current conversion number, and ensures that the medium and long-term value of traffic can be reflected in real-time decision-making.

[0074] For example, within 5 hours of exposure to the target traffic, the actual number of conversions, that is, the current conversion number, is 10 times, and the bonus coefficient is 1.6. Then the bonus conversion number is 10×1.6=16 times. This value reflects the possible conversion situation of the target traffic within the full 24-hour cycle.

[0075] In step 105, the server summarizes the additive conversion numbers of all flows in the current flow set to obtain the sum of the additive conversion numbers of all flows; at the same time, the main output header results of all flows are summarized to obtain the sum of the main output header results. A preset calibration method, such as rank-preserving regression calibration, linear regression calibration or quantile calibration, is used to adjust the sum of the main output header results so that the adjusted sum of the main output header results matches the sum of the additive conversion numbers. The result obtained after adjustment is the traffic decision estimate at the current moment. This process combines the calibration results of individual flows with the distribution characteristics of the overall flow, and achieves a balance between individual accuracy and overall consistency through sum matching, which not only retains the true potential reflected by the calibration of a single flow, but also corrects the estimation deviation at the overall level. The final output traffic decision estimate can dynamically adapt to fluctuations in user behavior, providing an accurate and practical basis for real-time decisions such as advertising ranking and resource allocation.

[0076] For example, the total additive conversion number of all flows in the current flow set is 1000 times, and the total main output header results are 800 times. The total main output header results are adjusted through the order-preserving regression calibration method to match the total additive conversion number of 1000 times. Then the estimated value of the flow decision at the current moment is 1000 times.

[0077] This application first obtains the full-cycle estimated conversion rate (main output header result) of the target traffic through online real-time reasoning to provide a long-term benchmark for calibration; at the same time, offline reasoning is performed on the traffic to generate a cumulative estimated conversion rate (sub-output header result) by time period and store it in the data pool. By matching the traffic generation attributes and exposure duration, the sub-output header results of the corresponding time period are matched from the data pool, and the additive coefficient is obtained by the ratio of the main output header result and the sub-output header result to quantify the degree of correlation between the full-cycle conversion and the current time period conversion. The current conversion number is calibrated using this coefficient to obtain the additive conversion number that integrates short-term actual conversion and long-term expectation. Finally, the sum of the additive conversion numbers of all traffic is summarized, and the traffic decision estimate is obtained by recalibrating it in combination with the sum of the main output header results. In this process, the offline reasoned time-segment data and the online real-time full-cycle data complement each other, which not only ensures the timeliness of calibration, but also improves the accuracy of the conversion estimate through multi-dimensional matching.

[0078] As an optional implementation method, if the main output header result cannot be matched to the corresponding sub-output header result due to special circumstances (such as data pool storage anomalies, or the time slice and model information of the target traffic are not covered by the data pool), the addition coefficient will be set to 1 by default. The core logic of this mechanism is that when there is a lack of time-segmented sub-output header results as a reference, by fixing the addition coefficient to 1, the current conversion number of the target traffic is kept at the original value during the calibration process (current conversion number × 1 = current conversion number), that is, the addition conversion number is equal to the current conversion number. The principle is to adopt a neutral calibration strategy in extreme scenarios where data matching fails, neither exaggerating nor depressing the value of the current conversion number, and avoiding the introduction of new deviations due to unfounded coefficient adjustments.

[0079] This setting of 1 ensures the robustness of the calibration process. Even if the sub-output header result is missing, the system can still complete the calculation of the additive conversion number through the default coefficient, ensuring that the generation of traffic decision estimates is not blocked. At the same time, a coefficient value of 1 means that the current conversion number directly represents its conversion expectation within the entire cycle. This conservative estimate not only complies with the principle of prudence when data is missing, but also reserves adjustment space for secondary calibration after subsequent data completion, avoiding the cumulative impact of temporary coefficient deviations on long-term decisions.

[0080] As an optional implementation, in step 102, the sub-output header results matching the main output header results within the corresponding time period are selected from the data pool according to the traffic generation attributes and exposure duration of the target traffic, including the following.

[0081] Step S11: selecting sub-output header results of multiple time periods that match the main output header result from the data pool according to the traffic generation attribute of the target traffic;

[0082] Step S12: Determine the corresponding time period of the exposure duration of the target traffic, where the corresponding time period is used to indicate that the traffic exposure duration is greater than m and less than n, where m and n are positive integers;

[0083] Step S13: Determine the sub-output header result of the target flow in the corresponding time period, wherein the sub-output header result is the cumulative estimated conversion rate from time 0 to time n.

[0084] In step S11, the server first analyzes the traffic generation attributes of the target traffic, including the exposure time slice (such as "09:15 - 09:20"), the name of the inference model (such as "Conversion Prediction Model B"), and the model version (such as "v2.1"). Based on these attributes, it retrieves the sub-output header results associated with the main output header results from the data pool. The data pool pre-stores the multi-period results generated by the offline inference of the target traffic, covering the estimated conversion rates for multiple cumulative periods such as 0 - 6h, 0 - 12h, 0 - 18h, 0 - 24h, etc. Through the precise matching of the traffic generation attributes, it is ensured that the selected sub-output header results and the main output header results (the full-cycle estimated conversion rate of online real-time inference) belong to the same target traffic and originate from the inference logic of the same model, avoiding data mixing across different traffic and models.

[0085] In step S12, the server calculates the time difference t from the exposure moment to the current moment of the target traffic, with the unit of hours, to clarify the exposure duration of this traffic. Based on the preset time segmentation rule, it compares t with each segmentation threshold to determine the corresponding time period it belongs to, which is presented in the form of greater than m and less than or equal to n (m and n are positive integers), such as 0 < t ≤ 6h, 6 < t ≤ 12h, etc. This step provides a basis for time dimension division for subsequent selection of appropriate sub-output header results by precisely quantifying the exposure duration and matching it to a specific time period. The core is to establish a one-to-one correspondence between the exposure duration and the time period, ensuring that each time node can find a matching segmentation standard, avoiding deviation in the selection of sub-output header results caused by ambiguous time division, and laying a foundation for subsequent precise calibration.

[0086] Exemplarily, the target traffic is exposed at 10:00, and the current time is 15:30. It can be calculated that the exposure duration t = 5.5 hours. Comparing with the preset segmentation rule, it is determined that this t belongs to the corresponding time period of 0 < t ≤ 6h.

[0087] In step S13, after the server determines the corresponding time period to which the target traffic belongs, according to the upper limit n of this time period, from the sub-output header data stored in the data pool, it selects the cumulative estimated conversion rate with a statistical range of 0 to n hours as the sub-output header result p. For example, when the corresponding time period is 0 < t ≤ 6h, it selects the cumulative estimated conversion rate from 0 to 6h; when the corresponding time period is 6 < t ≤ 12h, it selects the cumulative estimated conversion rate from 0 to 12h, and so on. This operation associates the upper limit of the corresponding time period with the statistical duration of the sub-output header result, ensuring that the selected sub-output header result can cover all the conversion expectations of this traffic from exposure to the current moment, and at the same time includes the conversion potential from the current moment to n hours. Its essence is to utilize the characteristic of cumulative estimation by time period to make the sub-output header result adapt to the current conversion progress of the traffic, avoiding calibration distortion caused by using results beyond the current time range or within too short a time range, and ensuring the accuracy of the subsequent calculation of the addition coefficient.

[0088] Exemplarily, it has been determined that the corresponding time period of the target traffic is 0 < t ≤ 6h, and the upper limit n of this time period is 6. Therefore, the cumulative estimated conversion rate of 3% for this traffic from 0 to 6h is selected from the data pool as the sub-output header result p.

[0089] This application calculates the exposure duration t and divides the corresponding time period, and then selects the cumulative estimated conversion rate with a duration of 0 to n hours according to the upper limit of the time period, enabling the sub-output header result to accurately match the conversion stage of the traffic at the current moment, avoiding missing the conversion expectations within this time period due to insufficient result coverage, and ensuring that the sub-output header result can truly reflect the conversion potential of the traffic at the current stage.

[0090] In addition, the sub-output header results for different time periods form a complete set of cumulative estimation systems. Results such as 0 to 6h and 0 to 12h progress step by step, covering all stages from exposure to the end of the full cycle. This design method enables the traffic to find corresponding reference standards at different exposure durations, making the calibration process transition naturally over time, avoiding the breakage of the calibration logic caused by jumps in time nodes, and enhancing the stability of the overall calibration process.

[0091] [[ID=第十二]]Finally, for high-latency conversion scenarios, selecting the cumulative estimated conversion rate with a duration of 0 to n hours as the sub-output header result includes both the conversions that have occurred and the delayed conversions that may occur but have not yet occurred within this time period. For example, when t = 8h (belonging to 6 < t ≤ 12h), the result from 0 to 12h has considered the conversions that may occur from 8 to 12h, enabling the subsequent addition coefficient calculated based on this result to reasonably reflect the impact of delayed conversions and avoiding the problem of underestimated estimation caused by traditional methods ignoring delayed conversions.

[0092] As an optional implementation, in step S11, based on the traffic generation attributes of the target traffic, selecting the sub-output header results of multiple time periods that match the main output header result from the data pool includes: based on the divided time slices of the target traffic and the model name and model version of the adopted inference model, selecting the sub-output header results of multiple time periods that match the main output header result from the data pool.

[0093] The server first analyzes the basic characteristics of the target traffic, extracting the time slice to which it belongs (such as the 5-minute time segment from 15:30 to 15:35, which is used to identify the specific time window in which the traffic was generated), as well as key information about the inference model used in the generation process of the traffic (including the model name, such as Conversion Prediction Model_E-commerce Edition; the model version, such as v4.2, which is used to distinguish the model parameters and inference logic of different iteration stages).

[0094] Based on this information, the server selects sub-output header results from the data pool that match the main output header results. The data pool pre-stores multi-period sub-output header results generated by offline inference for each flow (such as the cumulative estimated conversion rate for 0-6 hours, 0-12 hours, and 0-24 hours). This screening process requires three matching conditions: the target flow's time slices are identical to the sub-output header results' time slices, the inference model names are identical, and the model versions are unified. This screening process ultimately selects multi-period sub-output header results from the data pool that belong to the same time window and are generated based on the same model logic. This ensures that these sub-output header results and the main output header result (the full-cycle estimated conversion rate from online real-time inference) are the same inference results for the target flow.

[0095] As an optional implementation, before selecting a sub-output header result that matches the main output header result within a corresponding time period from the data pool based on the traffic generation attribute and exposure duration of the target traffic, the method further includes:

[0096] Step S21: Export the traffic after online inference is completed by time slices, and divide the traffic according to the model name and model version of the inference model used by the traffic;

[0097] Step S22: Perform offline reasoning on the divided traffic according to the traffic usage inference model, and output sub-output header results for each time period;

[0098] Step S23: storing the sub-output header results of the time periods into the data pool.

[0099] In step S21, after the online traffic is exposed, the system will cache the features used for model requests and store them in a table. When the offline estimation task exports the original feature data from the data pool, due to the large amount of data, time slices (such as 5 minutes) are used for sharded export to ensure the efficiency of data processing. At the same time, since there are multiple inference models running in parallel online, the exported traffic data needs to be divided according to the name and version of the inference model actually used for the traffic, and the divided data must be stored separately in a database, such as HDFS (Hadoop Distributed File System). This step reduces the scale of single data processing by time slice sharding, avoiding processing delays caused by excessive data volume; division by model name and version ensures the accuracy of data and model matching during subsequent offline inference, laying a data foundation for the generation of sub-output header results for time periods.

[0100] In step S22, the scheduling task will monitor the completion status of the data cached on the database in real time, and schedule the corresponding model version to perform offline inference on the divided traffic data based on the currently cached online estimated version. During the inference process, the model outputs the sub-output header results for each time period for each traffic (such as the cumulative estimated conversion rate for time periods such as 0-6h, 0-12h, etc.), and caches these results to HDFS. This step ensures the consistency of offline inference logic and online real-time inference by dynamically scheduling an offline model that is consistent with the online model version, avoiding deviations in sub-output header results caused by differences in model versions; at the same time, the design of time-based output results meets the subsequent needs of accurately selecting corresponding time period data based on different exposure times.

[0101] In step S23, the drop-table task continuously tracks the completion of offline inference. Once the time-segmented sub-output header results are cached in the database, they are written from the database to the data pool, forming a data pool for downstream calibration tasks to access. This process implements systematic storage and management of offline inference results, enabling downstream steps (such as the matching sub-output header results in step 102) to quickly retrieve the required time-segmented results from the data pool, reducing data access latency and improving the efficiency of the overall calibration process.

[0102] In this application, by dividing traffic data by time slice, model name, and version, and scheduling the corresponding version model for offline inference, we ensure that the offline-generated sub-output header results and the online main output header results are derived from the same set of features and inference logic, thus avoiding calibration deviations caused by model mismatches at the source of the data. In addition, time slice sharding exports solve the efficiency issues of large data processing, and the collaboration between database storage and table drop tasks enables efficient caching and call of inference results, allowing downstream calibration steps to quickly obtain the required data and support the timeliness requirements of real-time decision-making.

[0103] As an optional implementation method, the traffic exported by the shards is different batches of traffic. Each batch of traffic executes the processes of data division, data inference and data storage in sequence. The process steps of different batches of traffic are cross-parallel at the same time point.

[0104] The traffic exported by sharding is divided into different batches according to time slices, and the data processing of each batch of traffic adopts a full-process mechanism: in the data partitioning stage, the system splits the online traffic according to 5-minute time slices, and completes the traffic classification based on the model name and version. Multiple batches of tasks are started synchronously and run independently; in the data inference stage, the offline inference task performs model calculations on the completed batches based on the status feedback of the previous stage, and outputs the sub-output header results for the time period; in the data storage stage, the results of the completed inference are written to the database in parallel by batches to form a structured data pool.

[0105] The traffic exported by the shards is divided into multiple independent batches, and the data processing process of different batches of traffic is parallel. When a batch of traffic is performing data partitioning, another batch of traffic may have entered the data inference stage, and the batch started earlier may have begun data storage. This parallel mode achieves precise coordination through a scheduling state mechanism: each batch immediately updates the status identifier after completing data partitioning, triggering the offline inference task of the corresponding model; after the inference task is completed, the status is updated to promote the immediate execution of the storage task, realizing a seamless connection between partitioning, inference, and storage. Ultimately, all batches of data are offline estimated at the granularity of time slice * model version, ensuring that every record in the data pool accurately corresponds to a specific time range and model logic.

[0106] As an optional implementation, in step 105, the traffic decision estimate value at the current moment is determined based on the sum of the additive conversion numbers of all traffic at the current moment and the main output header result of each traffic, including the following contents.

[0107] Step S31: Calculate the sum of the additive conversion numbers of all flows at the current moment and the difference between the sum of the main output head results of each flow;

[0108] Step S32: Based on the difference, the sum of the main output head results of each flow rate is adjusted by a preset calibration method so that the sum of the adjusted main output head results matches the sum of the additive conversion numbers of all flow rates at the current moment;

[0109] Step S33: The sum of the main output header results after integration is used as the traffic decision estimate at the current moment.

[0110] In step S31, the server counts the additive conversion numbers of all flows in the current flow set, adds up these additive conversion numbers to obtain a total, recorded as total A; at the same time, the main output header results of each flow in the flow set are summarized to obtain total B. The difference between total A and total B is calculated. The difference can be the difference between the two (AB) or a ratio (A / B), so as to quantify the degree of deviation between the sum of additive conversion numbers and the sum of main output header results. This step provides a clear quantitative basis for subsequent adjustments. By accurately calculating the difference, the gap between the sum of the main output header results and the sum of additive conversion numbers that actually reflects the conversion potential of the entire flow cycle can be clearly known, which is the prerequisite for achieving accurate calibration.

[0111] In step S32, the server selects a preset calibration method (such as linear regression calibration, quantile calibration, dynamic weight calibration, Bayesian calibration, etc.) to adjust the sum B of the main output header results of each flow based on the difference calculated in step S31. Different calibration methods work for different data characteristics and scenarios. For example, when the sum A and the sum B are in an approximately linear relationship, linear regression calibration is used to construct a linear model to adjust B; when there are extreme values ​​in the data, quantile calibration is used to make separate adjustments within each quantile interval. The goal of the adjustment is to make the adjusted sum B' of the main output header results match the sum A of the additive conversion numbers of all flows at the current moment, that is, to make B' numerically consistent with A or within an acceptable error range. This step corrects the deviation between the sum of the main output header results and the sum of the additive conversion numbers through a scientific calibration method, so that the adjusted sum is more in line with the actual conversion potential of the flow.

[0112] In step S33, after the adjustments in step S32, the adjusted sum of the main output header results, B', matching the sum of the added conversions, A, is obtained. B' is directly used as the traffic decision estimate for the current moment. This estimate combines the main output header results from online real-time inference, the sub-output header results from offline inference, and the current actual conversion situation. It comprehensively reflects the conversion expectations for all traffic at the current moment over the entire cycle, providing a key quantitative indicator for real-time decisions such as advertising strategy formulation and resource allocation.

[0113] This application accurately calculates the differences and clarifies the degree of deviation between the sum of the main output head results and the sum of the added conversion numbers. It then uses a suitable preset calibration method to make adjustments and effectively correct the deviation, so that the final traffic decision estimate can more truly reflect the conversion potential of all traffic at the current moment, reducing the possibility of decision-making errors due to data deviations.

[0114] Figure 2 The complete flow chart of the method for determining the multi-period traffic estimation value by combining offline and online is shown in the figure. Figure 2 As shown, the process includes the following steps.

[0115] 1. Data preparation stage.

[0116] 1.1. Retrieve information about n traffic flows from the online data log Kafka. After processing, perform online real-time inference to generate the main output header results, denoted as {p1, p2, …, pn}, where pi represents the full-cycle estimated conversion rate of the i-th traffic flow. These main output header results are then labeled and stored in a table.

[0117] 1.2. After online exposure, cache the features of each traffic used for model requests and store them in a table. Offline tasks export the original feature data from the data table in 5-minute time slices, divide them into m batches according to model name and version, schedule offline inference of the corresponding version model, and output the sub-output header results of the time period {q it} (i is the traffic sequence number, t is the time period, covering 0-6h, 6-12h, 12-18h, 18-24h, etc.) and stored in the data pool.

[0118] 2. Matching and calibration calculation stage.

[0119] 2.1. Match the main output header result p by time slice, model name and version i With the sub-output header result q i .

[0120] 2.2. For each flow i, calculate the time difference t from its exposure to the current time i (hours), according to t i The range selection corresponds to the sub-output header result: 0 <t i ≤6h, use 0-6h conversion rate, 6 <t i ≤12h, use 0-12h conversion rate, 12 <t i ≤18h, use 0-18h conversion rate, 18 <t i When the conversion rate is less than 24 hours, the conversion rate of 0-24 hours is used and recorded as q it .

[0121] 2.3 If there is no match, set the addition coefficient k i =1, otherwise k i =p i / q it .

[0122] 2.3. Count the number of conversions ni that have been recovered for each flow, and calculate its added conversion number a i =n i ×k i , we get the sum A=∑a i , the sum of the main output head results B = ∑p i .

[0123] 3. Decision estimate determination stage.

[0124] 3.1. Calculate the difference between A and B, D = AB (or D = A / B), and adjust B to B' using a pre-set calibration method so that B' matches A.

[0125] 3.2. Take B' as the estimated value of the traffic decision at the current moment.

[0126] The beneficial effects that can be achieved by this application are as follows.

[0127] 1. Accurately address delayed conversions and improve the accuracy of conversion number estimates: By fully utilizing the time-segmented results of sub-output heads (such as 0-6h, 0-12h, etc.), targeted calibration is achieved in delayed conversion scenarios. For traffic that has not yet converted immediately after exposure but has the potential for delayed conversion, the sub-output head results of the corresponding time period are selected based on the time difference t from exposure to the current time to calculate the bonus coefficient, and the current recovered conversion number is associated with the long-term conversion expectation, so that the bonus conversion number can reasonably reflect the impact of delayed conversion. This mechanism effectively solves the problem of underestimation caused by traditional methods that rely solely on immediate conversion data, especially in business scenarios with long conversion cycles, improving the accuracy of conversion number estimates.

[0128] 2. Dynamically adapt exposure duration to cover the value of short-term unconverted traffic: The system fully considers traffic with short exposure time (e.g., t≤6h) that has not yet converted. By matching the results of 0-6h sub-output headers, it assigns a reasonable bonus coefficient to this type of traffic. Even if the current conversion number is 0, if the sub-output header results indicate a potential conversion probability during that period, the bonus coefficient can still reflect its long-term value based on the ratio of the main output header to the sub-output header. This avoids ignoring the potential value of traffic due to short-term non-conversion and ensures that the value of all traffic is included in the decision-making estimate.

[0129] 3. Combining offline and online operations to ensure timeliness and calibration efficiency: This integrated online and offline system architecture ensures the timeliness of online real-time reasoning (rapidly generating main output header results to support real-time decision-making). It also processes data in 5-minute time slices through offline tasks, schedules parallel inference for corresponding model versions, and rapidly generates sub-output header results and stores them in the data pool. This model not only meets the online demand for immediate response to traffic changes, but also provides detailed, time-segmented data support for conversion rate calibration through efficient offline computing. This achieves the synergy between real-time decision-making and precise calibration, significantly improving the efficiency and reliability of generating traffic decision estimates.

[0130] This application also provides an offline and online combined multi-period flow rate estimation determination device, such as Figure 3 As shown, the device includes:

[0131] An acquisition module 301 is configured to acquire a primary output header result of a target flow in a flow set, wherein the primary output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time inference;

[0132] Selection module 302 is configured to select, from the data pool, a sub-output header result that matches the main output header result within a corresponding time period based on the traffic generation attributes and exposure duration of the target traffic. The sub-output header result is the estimated conversion rate of the target traffic obtained through offline inference. The data pool stores the offline inference results of each traffic flow in different time periods.

[0133] A first determination module 303 is configured to determine a bonus coefficient based on the main output header result and the sub-output header result, wherein the bonus coefficient is used to indicate the degree of correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period;

[0134] A calibration module 304 is configured to calibrate the current conversion number of the target flow based on the addition coefficient to obtain the addition conversion number of the target flow;

[0135] The second determination module 305 is used to determine the traffic decision estimated value at the current moment according to the sum of the added conversion numbers of all traffic at the current moment and the main output header result of each traffic.

[0136] Optionally, the selection module 302 is used to:

[0137] According to the traffic generation attribute of the target traffic, the sub-output header results of multiple time periods that match the main output header results are selected from the data pool;

[0138] Determine the corresponding time period of the exposure duration of the target traffic, where the corresponding time period is used to indicate that the traffic exposure duration is greater than m and less than n, where m and n are positive integers;

[0139] Determine the sub-output header result of the target flow in the corresponding time period, where the sub-output header result is the cumulative estimated conversion rate from 0 to n time periods.

[0140] Optionally, the selection module 302 is specifically configured to:

[0141] According to the time slices of the target traffic and the model name and model version of the adopted inference model, the sub-output header results of multiple time periods that match the main output header results are selected from the data pool.

[0142] Optionally, the device is further used to:

[0143] Export the traffic after online inference is completed by time slices, and divide the traffic by the model name and model version of the inference model used by the traffic;

[0144] Perform offline inference on the divided traffic according to the traffic usage inference model, and output sub-output header results for multiple time periods;

[0145] Store the sub-output header results of multiple periods in the data pool.

[0146] Optionally, the traffic exported by the shards is different batches of traffic, and each batch of traffic sequentially executes the processes of data partitioning, data inference, and data storage. The process steps of different batches of traffic are cross-parallel at the same time point.

[0147] Optionally, the device is further used to:

[0148] If the main output header result cannot match the sub-output header result, set the bonus factor to 1.

[0149] Optionally, the second determining module 305 is configured to:

[0150] Calculate the sum of the additive conversion numbers of all flows at the current moment and the difference between the sum of the main output head results of each flow;

[0151] Based on the difference, the sum of the main output head results of each flow rate is adjusted by a preset calibration method so that the sum of the adjusted main output head results matches the sum of the additive conversion numbers of all flow rates at the current moment;

[0152] The sum of the main output header results after integration is used as the estimated flow decision value at the current moment.

[0153] like Figure 4 As shown, an embodiment of the present application provides an electronic device, including a processor 401, a communication interface 402, a memory 403 and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0154] The memory 403 is used to store computer programs.

[0155] In one embodiment of the present application, the processor 401 is configured to implement the offline and online combined multi-period flow estimate determination method provided by any one of the aforementioned method embodiments when executing the program stored in the memory 403 .

[0156] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the offline and online combined multi-period flow estimate determination method provided in any of the aforementioned method embodiments are implemented.

[0157] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0158] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, or of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiment.

[0159] It should be understood that the terms used herein are for the purpose of describing specific example embodiments only and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "comprise", "include", "contain" and "have" are inclusive and therefore specify the presence of stated features, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not to be construed as necessarily requiring them to be performed in the specific order described or illustrated, unless the order of execution is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0160] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. A method for determining multi-period traffic estimation values ​​by combining offline and online methods, characterized in that: The method comprises: Obtaining a main output header result of a target flow in the flow set, wherein the main output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time reasoning; Selecting a sub-output header result that matches the main output header result within a corresponding time period from the data pool based on the traffic generation attributes and exposure duration of the target traffic, wherein the sub-output header result is the estimated conversion rate of the target traffic obtained through offline inference. The data pool stores the offline inference results of each traffic flow in different time periods; Determining an addition coefficient based on the main output head result and the sub-output head result, wherein the addition coefficient is used to indicate the degree of correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period; Calibrate the current conversion number of the target flow based on the addition coefficient to obtain the addition conversion number of the target flow; The estimated value of the traffic decision at the current moment is determined based on the sum of the added conversion numbers of all traffic at the current moment and the main output head results of each traffic.

2. The method according to claim 1, characterized in that According to the traffic generation attribute and exposure duration of the target traffic, the sub-output header results matching the main output header results within the corresponding time period are selected from the data pool, including: According to the traffic generation attribute of the target traffic, selecting sub-output header results of multiple time periods that match the main output header result from the data pool; Determine a corresponding time period of the exposure duration of the target traffic, wherein the corresponding time period is used to indicate that the traffic exposure duration is greater than m and less than n, where m and n are positive integers; Determine a sub-output header result of the target flow in the corresponding time period, wherein the sub-output header result is a cumulative estimated conversion rate within a time period from 0 to n.

3. The method according to claim 2, characterized in that According to the traffic generation attribute of the target traffic, selecting the sub-output header results of multiple time periods that match the main output header result from the data pool includes: According to the divided time slices of the target traffic and the model name and model version of the adopted inference model, sub-output header results of multiple time periods that match the main output header result are selected from the data pool.

4. The method according to claim 1, wherein Before selecting, from the data pool according to the traffic generation attribute and exposure duration of the target traffic, a sub-output header result that matches the main output header result within a corresponding time period, the method further includes: Export the traffic after online inference is completed by time slices, and divide the traffic by the model name and model version of the inference model used by the traffic; Perform offline inference on the divided traffic according to the traffic usage inference model, and output sub-output header results for multiple time periods; The sub-output header results of the multiple time periods are stored in a data pool.

5. The method according to claim 4, characterized in that The traffic exported by the shards is divided into different batches of traffic. Each batch of traffic executes the processes of data partitioning, data inference, and data storage in sequence. The process steps of different batches of traffic are cross-parallel at the same time point.

6. The method according to claim 1, wherein The method further comprises: If the main output header result cannot be matched to the sub-output header result, the addition coefficient is set to 1.

7. The method according to claim 1, characterized in that Based on the sum of the added conversion numbers of all flows at the current moment and the main output header results of each flow, the estimated flow decision value at the current moment is determined, including: Calculate the sum of the additive conversion numbers of all flows at the current moment and the difference between the sum of the main output head results of each flow; Based on the difference, adjusting the sum of the main output head results of each flow rate by a preset calibration method so that the sum of the adjusted main output head results matches the sum of the additive conversion numbers of all flow rates at the current moment; The sum of the main output header results after integration is used as the estimated flow decision value at the current moment.

8. An offline and online combined multi-period flow rate estimation determination device, characterized in that: The device comprises: An acquisition module is used to obtain a main output header result of a target flow in a flow set, wherein the main output header result is a full-cycle estimated conversion rate of the target flow obtained through online real-time reasoning; a selection module, configured to select, from a data pool, a sub-output header result that matches the main output header result within a corresponding time period based on the traffic generation attributes and exposure duration of the target traffic, wherein the sub-output header result is an estimated conversion rate of the target traffic obtained through offline inference, and the data pool stores offline inference results for each traffic flow in different time periods; A first determining module is configured to determine a bonus coefficient based on the main output header result and the sub-output header result, wherein the bonus coefficient is used to indicate a correlation between the full-cycle conversion rate of the target flow and the conversion rate of the corresponding time period; a calibration module, configured to calibrate the current conversion number of the target flow based on the addition coefficient to obtain the addition conversion number of the target flow; The second determination module is used to determine the traffic decision estimated value at the current moment according to the sum of the added conversion numbers of all flows at the current moment and the main output header result of each flow.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the method according to any one of claims 1 to 7 when executing a program stored in a memory.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images